Mountpoint for Amazon S3
Open-source FUSE-style file client from AWS that mounts an S3 bucket as a local POSIX filesystem on a compute instance. Built on the **AWS Common Runtime (CRT)** library for high-throughput sequential access. Supports sequential and random *reads* + sequential *writes* (creating new files) — explicitly does NOT support arbitrary POSIX semantics like random writes, file truncation, or in-place modification. Fail-fast by design when unsupported operations are attempted, so applications hit clear errors rather than silently incurring expensive workarounds.
Definition
Open-source FUSE-style file client from AWS that mounts an S3 bucket as a local POSIX filesystem on a compute instance. Built on the **AWS Common Runtime (CRT)** library for high-throughput sequential access. Supports sequential and random *reads* + sequential *writes* (creating new files) — explicitly does NOT support arbitrary POSIX semantics like random writes, file truncation, or in-place modification. Fail-fast by design when unsupported operations are attempted, so applications hit clear errors rather than silently incurring expensive workarounds.
Production AI training and analytics workloads frequently read S3-stored datasets through filesystem-style APIs even though the underlying access pattern is "stream the whole file once" — but the dominant filesystem-style client `s3fs-fuse` tries to emulate full POSIX semantics on top of S3's object semantics, which produces surprising latency tails and silent multi-PUT-per-write storms. Mountpoint inverts that bet: don't try to be a real filesystem, just be the fastest possible thin layer for the read-mostly + sequential-write pattern. The result is **~6-8× the throughput of s3fs-fuse** on sequential workloads, at the cost of intentionally not supporting random writes or in-place mutation.
AI/ML training pipelines streaming dataset shards from S3, sequential-read analytics jobs (parquet/iceberg scans via Spark/Trino), high-throughput single-file uploads from compute instances, batch-data loading where the file-system API is required by the framework, and as the AWS-recommended alternative to s3fs-fuse / Goofys for read-heavy workloads.
Recent developments
- 6-8× the performance of s3fs-fuse on sequential reads + writes. Mountpoint significantly outperforms s3fs-fuse on the sequential-access pattern it's optimized for. Per Mountpoint inside-story — AWS Storage Blog.
- Repositioned post-S3-Files-launch (April 2026). With Amazon S3 Files GA on April 7, 2026, Mountpoint isn't dead — it's being repositioned for large-file throughput workloads where unsupported operations fail-fast by design, while S3 Files targets workloads needing full POSIX semantics. Per The Register — AWS S3 Files stress test.
- Built on AWS Common Runtime for high-throughput access. Mountpoint builds on the AWS CRT library which is purpose-built for high performance + low-resource usage on AWS endpoints. Per AWS — Mountpoint product page.
- Compared head-to-head with S3 Files + s3fs-fuse. The 2026 trio S3 Files / Mountpoint / s3fs-fuse splits the design space — S3 Files = managed full-POSIX, Mountpoint = open-source high-throughput read-mostly, s3fs-fuse = open-source full-POSIX-emulation with throughput tradeoffs. Per ComputingForGeeks — S3 Files vs Mountpoint vs s3fs.
- Goofys + s3fs-fuse + Mountpoint comparative analysis. Independent benchmarks show Mountpoint as the throughput leader for sequential AI/ML workloads, with s3fs-fuse still the choice when POSIX-semantics emulation is required at the cost of throughput. Per Medium — Maksym Lutskyi comparative analysis.